A river water quality inspection method based on unmanned ship calibration of unmanned aerial vehicle inversion model
By collaborating with unmanned surface vessels (USVs) and drones, a dynamic calibration algorithm and a dual-stream attention network were constructed, solving the problems of low efficiency and limited accuracy in existing technologies for river water quality inspection. This enabled rapid and high-precision water quality monitoring, making it suitable for complex water areas.
Patent Information
- Application Number
- CN202611127887.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-28
- Publication Date
- 2026-11-10
AI Technical Summary
Existing methods for inspecting river water quality are inefficient and make it difficult to achieve rapid, large-scale water quality monitoring. Furthermore, the accuracy of inversion models based on hyperspectral remote sensing technology is limited by static calibration and cannot cope with dynamic environmental changes.
The unmanned surface vessel (USV) is used to calibrate the UAV inversion model. Pollutant data is collected in real time by the USV and transmitted to the ground center. Combined with UAV hyperspectral imaging, a dynamic calibration algorithm is constructed. A dual-stream attention network is used to update the model parameters online, realizing the nonlinear mapping between pollutants and spectral features.
It enables rapid, large-area river water quality inspection, reduces manual labor, is suitable for complex water areas, and has high inversion accuracy, can calibrate errors in real time, shortens the source tracing cycle, and maintains high precision.
Smart Images

Figure CN123065982A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of water pollution source tracing technology, and in particular to a method for river water quality inspection based on unmanned vessel calibration of UAV inversion model. Background Technology
[0002] Currently, the mainstream methods for inspecting river water quality and tracing pollution sources mainly rely on traditional methods and rudimentary remote sensing technology. Firstly, it relies on manual inspection and laboratory analysis. This involves staff using inspection boats to manually collect samples at predetermined locations, then sending the samples to a laboratory for analysis. This method is not only time-consuming, resulting in pollution tracing cycles of several days or even weeks, but also inefficient. Furthermore, because it involves discrete point sampling, it fails to comprehensively reflect the overall spatial distribution of water quality in the river. Secondly, it employs hyperspectral unmanned aerial vehicle (UAV) remote sensing monitoring. While this method can achieve rapid scanning over large areas, its core water quality parameter inversion model typically relies on static calibration, i.e., a one-time correction using data from a small number of fixed ground sampling points before or after flight. This static model struggles to cope with dynamically changing atmospheric conditions (such as cloud cover and lighting changes) and aquatic environments (such as suspended solids concentration and wave disturbances) during flight, leading to limited inversion accuracy and susceptibility to errors. Third, although there are joint application cases of drones and unmanned ships, they are mostly concentrated on simple path tracking and image stitching. They have failed to deeply integrate the real-time, dynamic in-situ detection data of unmanned ships for online continuous calibration of hyperspectral inversion models, and thus cannot build an intelligent inversion system that can self-correct and adapt to the environment.
[0003] Therefore, the existing system has significant bottlenecks in terms of efficiency, cost applicability, and monitoring accuracy. Summary of the Invention
[0004] The present invention aims to solve the above-mentioned technical problems by providing a method for river water quality inspection based on unmanned vessel calibration of UAV inversion model.
[0005] The technical solution of this invention is a method for river water quality inspection based on unmanned surface vessel calibration of unmanned aerial vehicle inversion model, comprising the following steps: The unmanned surface vessel conducts mobile sampling along a preset route, collects pollution factor data in real time, and transmits the data to the ground control center via a wireless communication module; The drone is equipped with a hyperspectral imager to perform aerial scanning of the river area, acquire hyperspectral image data, and perform preprocessing and feature extraction. Based on real-time pollutant data transmitted by unmanned vessels, the UAV hyperspectral inversion model is dynamically calibrated to construct the mapping relationship between pollutants and spectral features, and output the inversion values of water quality parameters. The dynamic calibration algorithm includes an online iterative update mechanism that corrects the inversion model parameters in real time based on unmanned surface vessel sampling data.
[0006] As one implementation method, the online iterative update mechanism is configured as follows: Based on the spatiotemporal matching data of unmanned ships and drones, a real-time calibration sample set is constructed and managed by a sliding window. When the number of new samples or the time span of the samples reaches a preset threshold, the model incremental update based on the samples in that window is triggered, and the inversion model parameters are corrected using the mini-batch gradient descent algorithm. The inversion model is a dual-stream attention network structure that processes hyperspectral feature sequences and environmental auxiliary parameters in parallel, and integrates features through a weighted attention mechanism to construct a nonlinear mapping relationship between pollutants and spectral features.
[0007] In one implementation, the inversion model includes a convolutional neural network and a long short-term memory network. The convolutional neural network processes hyperspectral feature sequences and extracts local spectral features, while the long short-term memory network processes environmental auxiliary parameters synchronously measured by the unmanned vessel. The local spectral features and the environmental auxiliary parameters are weighted and integrated through an attention mechanism to construct a nonlinear mapping relationship between pollutants and spectral features.
[0008] As one implementation method, the unmanned vessel mobile sampling includes real-time detection of at least one pollutant among pH, ammonia nitrogen, COD, and turbidity, with a sampling frequency of not less than once per minute.
[0009] As one implementation, the preprocessing of the hyperspectral image data includes radiometric calibration, atmospheric correction, and image denoising, and the feature extraction of the hyperspectral image data includes at least one of spectral angle mapping, envelope removal, or first-order differential processing.
[0010] As one implementation, the dynamic calibration algorithm also includes an error compensation mechanism.
[0011] As one implementation, the error compensation mechanism is configured to estimate and compensate for water flow disturbances and illumination changes using a dual-wavelength compensation method.
[0012] As one implementation, the error compensation mechanism is configured to estimate and compensate for water flow disturbances and illumination changes through an extended state observer.
[0013] In one implementation, the unmanned vessel and the drone adopt a cooperative control strategy, and the preset route of the unmanned vessel is dynamically adjusted according to the hyperspectral anomaly region identified by the drone.
[0014] As one implementation method, the unmanned vessel and the drone achieve real-time data transmission via 5G communication.
[0015] The advantages of this invention compared to the prior art are reflected in the following three aspects: First, it enables large-scale, rapid inspections, shortening the traceability cycle. Traditional methods require fixed-point inspection by patrol boats, manual sampling, and results taking several days to be obtained in the laboratory, which is inefficient. The river water quality inspection method provided by this invention uses high-speed drones for area scanning, replacing slow, point-based manual sampling. Unmanned surface vessel navigation provides real-time ground truth values, replacing time-consuming laboratory analysis; cloud-based models automatically process and calibrate data, replacing manual data analysis. The combined efficiency improvements of these three stages significantly shorten the source tracing cycle.
[0016] II. Reduces labor input and equipment wear and tear, making it suitable for complex water areas. Unmanned surface vessels (USVs) and drones have replaced manually operated vessels, significantly reducing labor costs. In complex waters with strong currents, hidden reefs, shallows, or high pollution risks, they further reduce safety risks. USV and drone platforms are flexible, compact, and easy to operate, making them naturally suited for complex aquatic environments.
[0017] Third, high inversion accuracy is achieved through real-time calibration and error correction. Hyperspectral inversion is significantly affected by atmospheric and water body interference, making the model prone to inaccuracies. The river water quality inspection method provided in this invention introduces real-time data streams from unmanned surface vessels (USVs) as anchor points. A sliding window and incremental update mechanism ensure that the model continuously perceives system errors in the current environment. When the model deviates due to interference, the latest received USV ground truth data immediately pulls back the model parameters through a gradient descent algorithm, recalibrating its output to the correct trajectory. This provides anti-interference and adaptive capabilities, maintaining high-precision inversion at all times—something static models cannot achieve. Attached Figure Description
[0018] Figure 1 A flowchart of a river water quality inspection method based on an unmanned vessel calibrating an unmanned aerial vehicle inversion model, provided for an embodiment of the present invention. Detailed Implementation
[0019] The above and other embodiments and advantages of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0020] In one implementation, such as Figure 1 As shown.
[0021] This embodiment provides a method for river water quality inspection based on calibrating an unmanned surface vessel (USV) to an unmanned aerial vehicle (UAV) inversion model. The method includes the following steps: the USV conducts sampling along a preset route, collecting pollutant data in real time, and transmitting the data to a ground control center via a wireless communication module; the UAV, equipped with a hyperspectral imager, performs flight scanning of the river area to acquire hyperspectral image data, and performs preprocessing and feature extraction; based on the real-time pollutant data transmitted by the USV, the UAV hyperspectral inversion model is dynamically calibrated to construct a mapping relationship between pollutants and spectral features, and outputs inversion values of water quality parameters; wherein the dynamic calibration algorithm includes an online iterative update mechanism to correct the inversion model parameters in real time based on the USV sampling data.
[0022] The online iterative update mechanism is configured as follows: based on the spatiotemporal matching data of unmanned surface vessels and drones, a real-time calibration sample set is constructed and managed by a sliding window; when the number of new samples or the sample time span reaches a preset threshold, the model incremental update based on the samples in the window is triggered, and the inversion model parameters are corrected by a mini-batch gradient descent algorithm; wherein, the inversion model is a dual-stream attention network structure, which processes hyperspectral feature sequences and environmental auxiliary parameters in parallel, and integrates features by weighting through an attention mechanism to construct a nonlinear mapping relationship between pollutants and spectral features.
[0023] The inversion model includes a convolutional neural network and a long short-term memory network. The convolutional neural network processes hyperspectral feature sequences and extracts local spectral features, while the long short-term memory network processes environmental auxiliary parameters synchronously measured by the unmanned vessel. The local spectral features and the environmental auxiliary parameters are weighted and integrated through an attention mechanism to construct a nonlinear mapping relationship between pollutants and spectral features.
[0024] In this embodiment, the unmanned surface vessel (USV) conducts sampling along a preset route, such as covering a river cross-section along a zigzag pattern. The vessel is equipped with multi-parameter water quality sensors to collect real-time data on pollutants including pH, ammonia nitrogen (NH3-N), chemical oxygen demand (COD), turbidity (TUR), and dissolved oxygen (DO), with a sampling frequency of once every 30 seconds. Simultaneously, the USV records the precise latitude and longitude coordinates and timestamp of each data point via a GPS / BDS module. All data is transmitted in real-time to the ground control center via a 5G communication module.
[0025] Simultaneously, the UAV, equipped with a hyperspectral imager, flew along a planned route at an altitude of approximately 100 meters above the water surface. The hyperspectral imager continuously acquired hyperspectral image data with high spectral resolution. The UAV used a positioning module to obtain the precise geographical location and capture time for each frame of imagery. The raw image data was transmitted to the ground station via a high-speed image transmission link for preprocessing, including: radiometric calibration: converting the DN values of the imagery into radiance values; atmospheric correction: eliminating the influence of atmospheric molecules and aerosols to obtain surface reflectance products; and feature extraction: extracting spectral features such as first-order differential transforms and spectral absorption indices from the corrected spectral data to highlight characteristic bands related to water quality parameters.
[0026] In this embodiment, dynamic calibration is the core of the entire method. Its purpose is to continuously optimize and correct the UAV hyperspectral inversion model using real-time measured values from the unmanned surface vessel (USV), enabling it to output more accurate water quality parameter inversion values, namely predicted values of concentrations such as COD and ammonia nitrogen. Spatiotemporal matching data refers to pairing the true values of water quality parameters measured by the USV at a specific point in time with the hyperspectral feature data acquired by the UAV within the same geographical location and time window. This process ensures that the data used for calibration comes from the same water body at the same time, forming the basis for all subsequent processing. The calibration sample set consists of several calibration samples continuously received and stored by the system. Each successfully paired sample (geographical location, timestamp, measured value of pollutant, hyperspectral feature vector) constitutes a calibration sample. The system maintains a first-in, first-out queue as a sliding window to store the N most recently generated calibration samples (N is set to 200 in this embodiment). The window ensures that the model always learns based on the latest and most timely data, thereby quickly adapting to the dynamic changes in the water body. A model update is triggered when the number of new samples accumulated in the sliding window reaches M (e.g., 50), or when the time span between the oldest and newest samples in the window exceeds one hour. After triggering the update, the algorithm randomly selects a small batch (e.g., 32) of samples from the sliding window and calculates the loss between the current model prediction and the measured values of the unmanned surface vessel. Then, it calculates the gradient of the loss function with respect to the model parameters through backpropagation and fine-tunes the model's weights and biases along the gradient descent direction. This process is called incremental training or online iteration, which allows the model to continuously learn new data features without forgetting old knowledge, thus achieving self-evolution.
[0027] In this embodiment, a dual-stream attention network learning model is used as the core inversion model. The first stream is the spectral stream, based on a convolutional neural network (CNN). A CNN is a deep learning model adept at extracting local and spatial features. Here, hyperspectral data is treated as a one-dimensional spectral curve signal, and a one-dimensional CNN is used to automatically learn the local spectral features related to water pollution factors, such as the shape and depth of absorption valleys and reflection peaks in specific bands. The second stream is the environmental stream, based on a long short-term memory (LSM) network. LSM is a recurrent neural network adept at processing sequential data and remembering long-term dependencies. Here, the input is a sequence of environmental auxiliary parameters simultaneously measured by the unmanned surface vessel. These parameters include water temperature, air temperature, light intensity, and flow velocity, which significantly affect the spectral response. For example, water temperature affects algal activity, thus affecting the spectral features related to chlorophyll a. The high-level features extracted from the two streams are not simply concatenated but input into an attention mechanism layer. The attention mechanism mimics human cognition, automatically focusing on the features most important to the current inversion task. For example, when retrieving water quality parameters, the attention mechanism might assign higher weights to features in the 500-600nm wavelength band extracted by the convolutional neural network; conversely, when there are drastic changes in illumination, it might assign higher weights to illumination intensity features extracted by the long short-term memory network. This dynamic and adaptive weighted integration method greatly enhances the model's ability to construct a nonlinear mapping relationship between pollutants and spectral features. This relationship is not a simple formula, but a complex, data-driven function expressed through a deep learning model, with spectral and environmental data as input and water quality parameter inversion values as output.
[0028] In this embodiment, based on the aforementioned river water quality inspection method, a dynamically calibrated model processes the hyperspectral data of the entire river channel to generate a spatial distribution map of the inverted values of water quality parameters for the entire river channel. This allows for the intuitive and rapid identification of pollution discharge outlets, pollution zone extent, and diffusion trends. Compared to traditional river water quality inspection methods, this method, based on an unmanned surface vessel calibrating an unmanned aerial vehicle (UAV) inversion model, offers the following three advantages.
[0029] First, it enables large-scale, rapid inspections, shortening the traceability cycle. Traditional methods require fixed-point inspection by patrol boats, manual sampling, and results taking several days to be obtained in the laboratory, which is inefficient. The river water quality inspection method provided by this invention uses high-speed drones for area scanning, replacing slow, point-based manual sampling. Unmanned surface vessel navigation provides real-time ground truth values, replacing time-consuming laboratory analysis; cloud-based models automatically process and calibrate data, replacing manual data analysis. The combined efficiency improvements of these three stages significantly shorten the source tracing cycle.
[0030] II. Reduces labor input and equipment wear and tear, making it suitable for complex water areas. Unmanned surface vessels (USVs) and drones have replaced manually operated vessels, significantly reducing labor costs. In complex waters with strong currents, hidden reefs, shallows, or high pollution risks, they further reduce safety risks. USV and drone platforms are flexible, compact, and easy to operate, making them naturally suited for complex aquatic environments.
[0031] Third, high inversion accuracy is achieved through real-time calibration and error correction. Hyperspectral inversion is significantly affected by atmospheric and water body interference, making the model prone to inaccuracies. The river water quality inspection method provided in this invention introduces real-time data streams from unmanned surface vessels (USVs) as anchor points. A sliding window and incremental update mechanism ensure that the model continuously perceives system errors in the current environment. When the model deviates due to interference, the latest received USV ground truth data immediately pulls back the model parameters through a gradient descent algorithm, recalibrating its output to the correct trajectory. This provides anti-interference and adaptive capabilities, maintaining high-precision inversion at all times—something static models cannot achieve.
[0032] In one implementation, the river water quality inspection method based on the unmanned vessel calibration drone inversion model includes unmanned vessel mobile sampling, which includes real-time detection of at least one pollutant among pH, ammonia nitrogen, COD, and turbidity, with a sampling frequency of not less than 1 time / minute.
[0033] In this embodiment, to achieve high-frequency real-time detection of pollutants, the unmanned surface vessel (USV) is equipped with a multi-parameter water quality sensor. This sensor integrates a pH electrode, an ammonium ion (NH4+) selective electrode, a COD ultraviolet absorption spectroscopy sensor, and an optical turbidity sensor, enabling simultaneous measurement of these four key pollutants. The sampling process is automatically controlled by the onboard industrial control computer. During navigation, the system is set to automatically trigger a sampling measurement cycle every 30 seconds by default: a water pump draws water samples from approximately 50 cm below the surface, which flow through the measurement pool. Each sensor sequentially completes its readings, and the measurement data (pH value, ammonia nitrogen concentration - mg / L, COD concentration - mg / L, turbidity - NTU) are packaged together with the latitude and longitude coordinates obtained by the GPS module and the UTC timestamp. Subsequently, the data is transmitted to the ground control center in real time via a 5G module. This high-frequency sampling ensures that the calibration data has extremely high spatiotemporal resolution, providing a sufficient and timely true data source for the USV's inversion model.
[0034] In one embodiment, the river water quality inspection method based on the unmanned vessel calibration UAV inversion model includes preprocessing of hyperspectral image data including radiometric calibration, atmospheric correction and image denoising, and feature extraction of the hyperspectral image data includes at least one of spectral angle mapping, envelope removal or first-order differential processing.
[0035] In this embodiment, the raw hyperspectral data acquired by the UAV undergoes a series of processing steps to extract useful information. Radiometric calibration: Using the radiometric calibration coefficients provided by the hyperspectral imager, the digital quantization values (DN values) of the image are converted into absolute radiance values. This step eliminates the sensor's own response differences. Atmospheric correction: Inputting meteorological data during flight (such as visibility and atmospheric patterns), the MODTRAN radiative transfer model is used for inversion to eliminate the effects of atmospheric molecule and aerosol scattering and absorption, ultimately obtaining the true surface reflectance data of ground objects. This is a crucial step in eliminating atmospheric interference. Image denoising: A filtering algorithm is used to smooth the spectral curve. This method effectively suppresses random noise while preserving the shape and width characteristics of the spectral signal to the maximum extent. After preprocessing, feature extraction is performed to highlight information related to water quality.
[0036] In this embodiment, first-order derivative processing is preferred, where the first derivative is calculated band by band for the smoothed reflectance spectral data. Derivative processing can eliminate background noise and enhance subtle changes in the spectral curve; the positions and depths of these inflection points are highly correlated with specific water quality parameters. As an alternative implementation, envelope removal is used. This method can be used if direct analysis of absorption characteristics is required. By calculating the ratio of the reflectance spectrum to its convex envelope point by point, the spectral data is normalized to the range [0,1], thereby highlighting absorption characteristics and performing spectral feature matching.
[0037] In one implementation, the river water quality inspection method based on the unmanned vessel calibrating the UAV inversion model employs a cooperative control strategy between the unmanned vessel and the UAV, and the preset route path of the unmanned vessel is dynamically adjusted according to the hyperspectral anomaly region identified by the UAV.
[0038] In this embodiment, the specific flow of the control strategy for the unmanned surface vessel (USV) and the unmanned aerial vehicle (UAV) is as follows: 1. The UAV flies along a preset global path and processes hyperspectral data in real time to generate a preliminary water quality distribution map. 2. The algorithm at the ground control center analyzes the distribution map in real time and automatically identifies areas with abnormal reflectance, such as suspected high-concentration pollution areas, whose spectral shapes are significantly different from the surrounding water bodies. 3. Once an abnormal area is identified, the algorithm immediately sends the central geographic coordinates and extent of the area to the USV. 4. After receiving the instruction, the USV's path planning module dynamically replans its current route. It first heads towards the abnormal area and automatically generates a dense "X"-shaped or grid-shaped fine sampling path within its range to update the original preset cruise path. 5. The USV conducts high-intensity sampling in this area, providing the UAV model with dense ground truth data for this key area, thereby achieving a closed loop and improving the efficiency and accuracy of the source tracing investigation.
[0039] The specific embodiments described above further illustrate the inventive purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, or improvements made by those skilled in the art within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for river water quality inspection based on unmanned surface vessel (USV) calibration of UAV inversion model, characterized in that, Includes the following steps: The unmanned surface vessel conducts mobile sampling along a preset route, collects pollution factor data in real time, and transmits the data to the ground control center via a wireless communication module; The drone is equipped with a hyperspectral imager to perform aerial scanning of the river area, acquire hyperspectral image data, and perform preprocessing and feature extraction. Based on real-time pollutant data transmitted by unmanned vessels, the UAV hyperspectral inversion model is dynamically calibrated to construct the mapping relationship between pollutants and spectral features, and output the inversion values of water quality parameters. The dynamic calibration algorithm includes an online iterative update mechanism that corrects the inversion model parameters in real time based on unmanned surface vessel sampling data.
2. The method for river water quality inspection based on unmanned surface vessel calibration of UAV inversion model as described in claim 1, characterized in that, The online iterative update mechanism is configured as follows: Based on the spatiotemporal matching data of unmanned ships and drones, a real-time calibration sample set is constructed and managed by a sliding window. When the number of new samples or the time span of the samples reaches a preset threshold, the model incremental update based on the samples in that window is triggered, and the inversion model parameters are corrected using the mini-batch gradient descent algorithm. The inversion model is a dual-stream attention network structure that processes hyperspectral feature sequences and environmental auxiliary parameters in parallel, and integrates features through a weighted attention mechanism to construct a nonlinear mapping relationship between pollutants and spectral features.
3. The method for river water quality inspection based on unmanned surface vessel calibration of UAV inversion model as described in claim 2, characterized in that, The inversion model includes a convolutional neural network and a long short-term memory network. The convolutional neural network processes hyperspectral feature sequences and extracts local spectral features, while the long short-term memory network processes environmental auxiliary parameters synchronously measured by the unmanned vessel. The local spectral features and the environmental auxiliary parameters are weighted and integrated through an attention mechanism to construct a nonlinear mapping relationship between pollutants and spectral features.
4. The method for river water quality inspection based on unmanned surface vessel calibration of UAV inversion model as described in claim 2, characterized in that, The unmanned vessel mobile sampling includes real-time detection of at least one pollutant among pH, ammonia nitrogen, COD, and turbidity, with a sampling frequency of no less than once per minute.
5. The method for river water quality inspection based on unmanned surface vessel calibration of UAV inversion model as described in claim 2, characterized in that, The preprocessing of the hyperspectral image data includes radiometric calibration, atmospheric correction, and image denoising. The feature extraction of the hyperspectral image data includes at least one of spectral angle mapping, envelope removal, or first-order differential processing.
6. The method for river water quality inspection based on unmanned surface vessel calibration of UAV inversion model as described in claim 1, characterized in that, The dynamic calibration algorithm also includes an error compensation mechanism.
7. The method for river water quality inspection based on unmanned surface vessel calibration of UAV inversion model as described in claim 1, characterized in that, The error compensation mechanism is configured to estimate and compensate for water flow disturbances and illumination changes using a dual-wavelength compensation method.
8. The method for river water quality inspection based on unmanned surface vessel calibration of UAV inversion model as described in claim 1, characterized in that, The error compensation mechanism is configured to estimate and compensate for water flow disturbances and illumination changes through an extended state observer.
9. The method for river water quality inspection based on unmanned surface vessel calibration of UAV inversion model as described in claim 1, characterized in that, The unmanned vessel and the drone employ a collaborative control strategy, and the preset route of the unmanned vessel is dynamically adjusted based on the hyperspectral anomaly regions identified by the drone.
10. The method for river water quality inspection based on unmanned surface vessel calibration of UAV inversion model according to claim 1, characterized in that, The unmanned vessel and the drone achieve real-time data transmission via 5G communication.
